Podcast – Agentic AI and the Future of Dementia Research

Voiceover

The Dementia Researcher Podcast, talking careers, research, conference highlights, and so much more.

Professor Louise Serpell

Hello and welcome to the Dementia Researcher Podcast. Today we have the man behind the Alzheimer's Insights AI Prize and the two teams who won it. We are talking artificial intelligence, agentic AI, and what they really mean for dementia research. Hello, I'm Professor Louise Serpell from the University of Sussex in the UK and I'm delighted to be hosting this week's show, where we're going to be exploring something that has been generating quite a bit of conversation across the field, which is the role that AI and in particular what is now being called agentic AI may play in dementia research.

So to set the scene, back in March, the Alzheimer's Disease Data Initiative announced the winners of its Alzheimer's Insights AI Prize. It was originally designed as a single $1 million prize, but the quality of the proposals were so high that the judges ended up naming two winners and doubling the pot. With me, I have three people in the room who can tell the story and tell us how AI can make a difference. Dr. Niranjan Bose, Interim Executive Director of the Alzheimer's Disease Data Initiative and Managing Director for Health and Life Sciences at Gates Ventures.

Niranjan, welcome. Or Bose, welcome.

Dr Niranjan Bose

Thank you, Louise. Thank you for having me here.

Professor Louise Serpell

I am also joined by Jonathan Hoover from Prima Mente, one of the winning teams with their platform Parthenon. Jonathan, hello.

Jonathan Hoover

Hi Louise. Thanks for having me.

Professor Louise Serpell

And finally, Kexin Huang representing Biomni AD, the other winning team, which is a collaboration between Stanford University and the Icahn School of Medicine at Mount Sinai. Kexin, welcome.

Kexin Huang

Thank you, Louise. Excited to be here.

Professor Louise Serpell

We're going to start with Bose to set the scene. Then we will hear from each one of our winners about what they're actually building and we'll finish with some broader reflections on AI in this field and some tips for any ECRs listening who fancy putting themselves forward for something like this in the future. Bose before we get into the prize itself, I think it's worth setting the scene. The AD Data Initiative is one of those organisations that people in the field have probably come across, but not everyone will know exactly what it does or how it came about. Can you start us there?

Dr Niranjan Bose

Absolutely, Louise. The Alzheimer's Disease Data Initiative as a journey for us started seven, eight years ago and driven fundamentally by the question that the field of Alzheimer's and related dementias has had some challenges. Late stage clinical failures, difficulty to diagnose the disease in a non invasive, scalable manner. So our principal Bill Gates asked the question, what if we were able to make the historical or the retrospective data sets more available, accessible and analyzable to the global community without really defining who that community is.

It could be anyone from any walk of life who might have an idea because maybe a family member, maybe a friend was affected by the disease and they wanna give back to society. So that was our fundamental premise. And five years ago, we launched this as a standalone nonprofit organisation with the one mission of making data more broadly available, accessible to researchers globally in the neurodegeneration field.

Professor Louise Serpell

Well that sounds fantastic and I just wonder if I can ask you a little bit about yourself and how you got to be in the position that you are now in. You told me earlier that you are not an AI person by training.

Dr Niranjan Bose

Yes, and I'll tell you one more thing. I'm not a neuroscience researcher either. So combine that and you may wonder why I am sitting here talking to you. I'm a microbiologist by training. I joined the Gates Foundation about 20 years ago and most of my work at the Gates Foundation was on vaccines, primarily paediatric, diarrheal and respiratory vaccines. Joined Gates Ventures, which is Bill Gates' private office, about a dozen years ago. And the primary role was and continues to be, to serve as a scientific advisor to Bill on healthcare and life sciences topics.

One of the topics that sparked Bill's curiosity and interest was Alzheimer's disease. So we embarked on a learning journey and built the learning agenda for him to understand the challenges, the bottlenecks, and perhaps the reasons why that might have led to late stage failures in the field. And long story short, here we are about eight years ago, Bill made his first philanthropic investment in Alzheimer's disease through Gates Ventures. And as I indicated earlier, five years ago we launched the Alzheimer's Disease Data Initiative.

And because of my role at Gates Ventures, I get to serve as the interim executive director currently.

Professor Louise Serpell

That sounds really interesting. So I've heard of the AD Workbench and I wondered if you could tell us a little bit about what that is and how people can use it.

Dr Niranjan Bose

Absolutely. If you think of making data sets available, you also need to make sure that you embrace interoperability and secure workspaces. Let me break both of those down. There are a lot of rich data sets that reside in global data platforms such as Dementia Platforms UK, Vivli, the Global Alzheimer's Association Interactive Network or GAAIN. How do you build on that rather than saying, "Hey, we're going to build the next best centralised data repository" and we didn't wanna do that.

That's why we picked interoperability. It's our number one technical objective. And AD Workbench is that interoperability layer that stitches together nearly dozen global platforms that house rich data sets from various geographies including US, Europe, Asia, and Latin America and the likes. The second part on secure workspaces is making data sets available is a great thing.

It's a starting point. You need to get on "first base" to get their home plate. Secure workspaces are the next part of that journey. For people who don't have access to compute or tools that allow them to combine these data sets to analyse these data sets in a secure manner while still adhering to the privacy and governance and permissioning rules of the data contributors, these workspaces are cloud based environments that are built within AD Workbench that democratise access to cloud, to data sets and compute.

Professor Louise Serpell

That sounds really exciting. And from my point of view as a dementia scientist, I think that's a really exciting perspective on how collaboration can actually be conducted in a really diverse way so that you get, give access to everybody, I suppose, who has computer access. So that sounds fantastic.

Dr Niranjan Bose

Absolutely. I'm reading the Infinity Machine now, so if I want a six year old player, chess player to really go ahead and build a model, we need to make sure we democratise access to the high school student, to the middle school student, to the college student.

Professor Louise Serpell

Yes, and from whatever country. So that sounds fantastic. Thank you. So how did this all end up with AI? in a way it makes sense. You've got a huge amount of data, from all of these different platforms. And can you tell our listeners exactly what agentic AI is?

Dr Niranjan Bose

Great question. I will attempt to answer that question for you. We are joined by two experts who perhaps are more adept at giving us a very clear answer. But as I understand it, and let me break it down into maybe a real life analogy here. We all use automation. We automate things, let's say a Google search to book an airline ticket. We are automating tasks. But when you think about an AI agent, think about it as having a little bit more autonomy. It can think for itself and it can give you answers that might go beyond do step one and step two.

But when you think about agentic AI, and let's say I want to plan a birthday party for my son, I can say plan a birthday party for me and he loves Premier League. The agentic AI solution should be able to look at my calendar, the dates the family's free, prepare an invitation list, find a venue, order the themes, order the napkins, order the cake that's all themed in his favourite thing, send it out and comes back to me and says, "Okay, all of that's done. Here's the number RSVPs you've received." So it's going beyond automation, it is going beyond one independent task.

It's about thinking it through end to end.

Professor Louise Serpell

Thank you that was really helpful for me 'cause I did look it up and your explanation was perfect. So if I'm right in understanding this, if I compare it to say the usual AI chatbots that people are used to using now, presumably this goes beyond one question or one sentence and continues that to its ultimate conclusion in actually completing the task.

Dr Niranjan Bose

It absolutely does. In the realm of data sets, now that we have about 350 or so data sets across these dozen interoperable platforms, if I'm a researcher and I come in and I don't know how to code, but I have a very, very good question to ask. And let's say the question is why do women get more Alzheimer's than men? And if I go to an agentic AI tool and I write it down and say, "I'm trying to answer this question, can you help me using the AD Workbench, find out how I should go about answering it?

Even better, why don't you put in requests for all those data sets, put it in my secure workspace, run your agents that can do the regression analysis and just give me the charts with five hypothesis that I could test in my lab? That is what we want these agentic AI solutions to do.

Professor Louise Serpell

Well, I would love to know the answer to that. I have a student who's actually working on that particular problem and that will be fantastic to know the answer that AI would come up with. So I understand that from what you've said and I'm just wondering why do you think data has been such a difficult problem specifically for Alzheimer's and dementia research specifically because it sounds the Gates Ventures have identified that particularly when obviously there are lots of other conditions and diseases that it could have been used for.

Dr Niranjan Bose

Another fantastic question. Everybody who funds, participates and conducts studies wants to make that data available for novel insights to be generated. So the coalition that came together of funders, academic researchers, industry partners to get the Alzheimer's disease data initiative started, realised that there are perhaps three key barriers that we'd have to overcome.

The first one I spoke to is interoperability. When data sits in siloed environments, it's really hard to bring them together and do the combined analysis. The question on why do women get more Alzheimer's than men, you might have to bring, let's say 10 data sets that live in four different platforms together. And that ability to do the combined analysis has been a challenge in the past. And I believe we built necessary bridges so we can still build more bridges. But I believe we have a good quorum on interoperability now.

The second aspect is permissioning and governance. We need to absolutely make sure we are adhering to the consent forms, the nature of the data sharing that was agreed upon when the study was conducted. If we could build agents that understood every aspect of what went into the consent form. And when I'm a user who's requesting a data set, can the agent actually do the review of my application and make it easier for the permissioning person to say, yeah, this person checks all your boxes.

You can go ahead and greenlight this approval without waiting three weeks or three months in a matter of maybe a day or two days. You reduce the friction on access. The third aspect I would say is some of these studies perhaps where data was collected in a pen and paper format, I'm talking studies maybe that were done 50 years ago, 60 years ago and they've not been digitised or they've not been harmonised.

For all of those, we need to bring in additional resources, move them into a digital format to help them with the curation or the harmonisation. And some of them may not have robust data dictionaries. I mean those are the gaps that we want to fill, whether manually or using tools, automation and agents. So we can reduce the friction on data sharing.

Data is an extremely valuable asset if used appropriately, if the right questions are posed to it. And that's the journey we're on.

Professor Louise Serpell

And presumably you are updating the data all the time. So if somebody, for example, had cohort data on Alzheimer's disease, would they contact you or to upload that or to to contribute that?

Dr Niranjan Bose

Absolutely, and it's a two way street. We reach out to cohorts and data contributors and principal investigators and many times they reach out to us as well because the intent and the wish to share the data to make it available is there. Many times it's understanding the constraints, the barriers and those legal considerations that sometime have to be unblocked or unpacked so that we could make it a lot easier for the data contributors to share the data and data does get updated, and that's why I think having the right version and having the right data dictionaries that go with the updated data sets are critically important as well.

Professor Louise Serpell

Yes, I was thinking of an example where at the ADPD meeting, Nova Nordisk gave several talks about their GLP1 trials and talked about how they were going to share the data, which I thought was fantastic because it feels quite new to me that researchers or companies would be sharing their data with everybody to have others analyse it. So that's presumably the thing that would be shared on your platform.

Dr Niranjan Bose

Absolutely. And even if it's shared on a different platform, we will strive to build the interoperability bridge. So for our user, the ability to access the data is as minimally frictionful as possible.

Professor Louise Serpell

Yeah. Wonderful. So we should turn to the prize. We have two winners here. Congratulations, Jonathan and Kexin. So I just wonder before we move on both, if we can talk about the Alzheimer's Insights AI prize, which was launched in August, 2025 and what the original intention behind it was and why did you feel that the competition was the right mechanism?

Dr Niranjan Bose

Another great question. The prize was thought about as an idea at the Alzheimer's Association's International Conference a year ago in 2025. And we were faced with two choices. One, we could commission our own set of AI agents in agentic AI solutions through a selection process of partners, maybe companies that could build it for us. And the other choice was to put out a million dollar prize and see what ideas came from the global community. And we were quickly convinced that option, the second option was the better way to go because what if we failed to think about the best idea or the better idea in our whiteboarding exercise?

So why narrow it down to just our conference room and our ideas? And that's why we chose to go with the prize. Now we were still faced with some critique in that when have ever prizes solved problems and it's a fair critique, but in this case we were able to overcome some of those critiques by saying we're not looking to solve the problem.

We're building tools that will enable users and other researchers to come in and solve some problems and ask them questions. So if we were to do it again, I think we'd do the prize all over again. It was a excellent experience. We got it done in about eight, nine months. The diversity, the ideas was striking. And we have two fantastic winners, thanks to our jury and the applicants.

Professor Louise Serpell

Were you surprised when you got 180 submissions?

Dr Niranjan Bose

I was surprised that it was not more than 180. The speed at which the field is moving, I would've expected to see a lot more. I'd have expected to see 500, and I think my team informed me nearly 500 folks registered on the portal to submit an application and we did that so we could gauge how many judges we need. And of the 500, only about 180 or 200 ended up finishing the application.

Professor Louise Serpell

That's an interesting statistic that probably reflects most grants and applications and prizes. I would think so. Although I might think that in the AI field, a million dollars is perhaps very little money. Ah, well we'll talk to the others about that. Thank you so much Bose. That was really useful grounding. Let's bring in one of the winning teams now. Kexin, your team's solution has been described as an AI co scientist for Alzheimer's research. So I'd to start with what that actually means in practise, because co scientist is a phrase that gets used a lot at the moment.

Can you walk us through what Biomni AD does and the language you'd use with a colleague who works on dementia but isn't an AI specialist?

Kexin Huang

Yes, thanks for the question. So I think a useful context is actually to look at how scientists or dementia researcher has been doing biomedical research today 'cause if we really look at the day to day life, a typical researcher will need to navigate through many different tools and software databases, literature, and then even working with other scientists from different expertise to get work done.

So it's a very manual complex process. So what the Biomni AD does, it's really think about how do we automate, how do we scale up this process from the question to result? So Biomni AD intrinsically is a, you can think of it as a virtual research intern. So instead of the scientists handle, perform all the tasks, now it's the agent to perform all the tasks.

And the scientists just need to take the steering wheel and ask the question, take the ideas, and then Biomni AD will handle all the taskss. So the key idea of Biomni, which is a precursor of the Biomni AD project, is a general purpose biomedical agent that can do all kinds of different biomedical research tasks ranging from ideations, design molecules, experiment protocols, and then even doing the data analysis and so on.

So our goal is really want to make a platform that is specialised for Alzheimer's disease and then it can help scientists to perform all kinds of day to day tasks in a more autonomous way so that it can save a lot of time. So scientists can also focus more on the discoveries in contrast to handle all the executions.

Professor Louise Serpell

It sounds amazing and transformative for the scientist's workload. Do you think that scientists will embrace it or do you think that they might feel that their autonomy or their work is being taken from them?

Kexin Huang

Yes, I think there's definitely will be a transition period, but I would say all the scientists have been experimenting with the platform. They all end up liking it because it just saves so much of their time so they can focus more on the ideas and the discovery, which is the fun part of the science in contrast to the more tedious part of the science. Yes, so I can imagine this is a great amplifier for the scientists in contrast to a replacement of a scientist.

Professor Louise Serpell

Yes, absolutely. And just from a personal point of view of having used AI and then decided, "Oh no, I don't the way that AI has done this." I'm wondering if you have the opportunity to change the pathway and put in your own input and say, "Oh no, I don't want to do the experiment this, I want to do it slightly differently." Do you have a interactive platform that allows you to do that?

Kexin Huang

Exactly so the interaction and the collaboration is the key features of Biomni AD because we understand scientists when they work day to day, it's also a iterative process. It's never one step and then they have result. Yes, so that's why we actually design the agent to encourage to proactively collaborate with the scientists, by asking the right questions, before you launch the task and make a plan to work with the scientist to align with the plan. So it's almost your real research interns, right, before it goes out for a task, it align with you first and then really iterate together to get work done.

So that's a key feature that we are emphasising and pushing forward a lot on.

Professor Louise Serpell

Fantastic. So it looks the Biomni platform was developed initially with Stanford and Mount Sinai, and I wonder if you can tell us a bit about that partnership?

Kexin Huang

Yes, exactly; so initially Biomni AD is a project through my PhD at Stanford. So it's an open source project and allow anyone is can contribute to it and initially Biomni is designed to be general purpose, so it can do across disease areas. And after the open sourcing, we have interest from the community to contribute and to specialise in different disease areas. So that's where Qwan from Mount Sinai who's a very experienced and well established researcher in Alzheimer's disease.

He's also excited about the future of agents in Alzheimer's disease and he reached out to us to say if we can partner together to specialise Biomni to Biomni AD. So he's going to still having the specialisation effort. And then we have been working together for a few months trying to bring Biomni to the Alzheimer's disease researchers system.

Professor Louise Serpell

I'm slightly intrigued about your background. So where have you come from to get to this stage? And in terms of whether you are a computer scientist or a biologist or what was your intention when you first started working on this project?

Kexin Huang

Yes, so I'm a computer scientist by training. So I've been working on this AI and biomedicine research for almost a decade. So I was intrigued by biology during my junior year in college, which is almost 10 years ago. And since then had just been working on AI and biomedicine yeah, ever since, working on all kinds of different problems in drug discovery from target discovery, human genetics, and all the way to molecule design, even clinical trial design and so on. So definitely the AD is also very personal for me.

I actually have, I think 10 years ago I did a 23andMe test. I also have the APOE variant. So that's why I'm also very passionate about this as well.

Professor Louise Serpell

Yes, many of us do. Yes. So can you tell me a little, just to maybe give an example of what a typical task looks when a researcher uses the tool?

Kexin Huang

Yes, so there's all kinds of tasks that a platform can enable. Maybe I can give an example on the data analysis side. So traditionally for example, you have a single cell RNA sequencing data set. You want to really understand what are the cell types annotated it. It often involves a very long process of, proper process data set through the quality control, through the PCA and through the UMAP. And also do a lot of literature research and manually look at each clusters and then try to identify what's marker gene and so on.

So traditionally, if the data set is very large, it can take a team of scientists to several months to get work done. But right now you can, in our platform, you can just upload the account matrix data set and then let the agent know I want to do cell annotation for this single cell RNA sequencing data sets. And then the agent will ask you some clarification questions or make a plan. And then after that, it will just going to do everything when I just mentioned autonomously and then return a report that talk about all the cell type annotations and also interpreted the data as well.

And then you can iterate with the agent to ask follow up questions and then to make real interesting hypothesis and discoveries.

Professor Louise Serpell

It really sounds that one might not need to think very much. Is that true, do you think?

Kexin Huang

Yes, I still think human is still under control. So how do you come up with the right question? How do you give feedback to the scientists, to the agent? It's still dependent on the thinking of the human scientist, especially the space of hypothesis is just so large and the humans still need to think about it and then pick the one that they want to pursue.

Professor Louise Serpell

Yes, so a young student perhaps couldn't say, "Oh, well I don't need to read any of the literature because the AI will just do it for me." I perhaps hope.

Kexin Huang

Yes, AI will help you summarise and then can make it faster. But I think every young student still need to have a solid and fundamental knowledge about a space because AI is also going to generate a lot of traces and tags and output. But at the end of the day, if you want to make this high stakes hypothesis and send it to the web lab for example, you still need humans to review and then the review process still requires fundamental knowledge understanding about this task. So I definitely will still argue that yeah, my scientist knowledge and the job is still very well retained.

Professor Louise Serpell

So the judges highlighted multimodal data spanning omic single cell biomarkers and clinical domains. So what's really strikes me about that is how much of that data is easily accessible, presumably that's from the database AD data initiatives and what does that mean for the kinds of questions that researchers can ask?

Kexin Huang

Yes, I think this is where the AD Workbench platform really shines because Biomni only without the AD data set, is a general purpose agent. So without this data set, I actually cannot ask, answer many meaningful questions for the AD researchers. So that's why our main effort is actually trying to integrate these AD Workbench data sets in a very native fashion to the Biomni platform such that agent can utilise the right data set with, or a combination of data set with any question. And that's relevant for the AD research.

Professor Louise Serpell

Okay, and I was also thinking that that must mean that, that there's much less overlap in terms of experimental work. So people aren't needing to do the same experiment more than once that they're just accessing the data that's already been uploaded rather than having to do that, that experiment work themselves.

Kexin Huang

Right, yeah, I think definitely it's a a tremendous resource of experiments and if it's already covered in the existing data set, it definitely does not need to be redone. But I'll also argue that there's still a tremendous amount of missing experimental data out there. Maybe the agent, the role is actually it can looking at the existing data set and come identify the gaps and identify the risk and also suggest what experiments are needed to answer the questions. Yeah but I'll still think experiment is extremely valuable to improve the Biomni AD platform.

Professor Louise Serpell

Okay, thank you very much. So what were the difficulties that you needed to overcome? You've been working on this a long time. What did you have to do to make sure that this actually works for its intended use?

Kexin Huang

Yes, I think there's definitely many challenges. I think right now building a prototype is relatively easy with the AD coding agent has been getting very popular. Yes, but when we really letting the platform to be used by scientists, just identify so much gaps, the agent need to be, have as close to zero hallucinations as possible. Agent had, needs to be scalable too, so a lot on the infrastructures as well. And they also integrated natively with the AD Workbench and with the right UI and UX.

So all of these are challenges. Yeah. And I usually make an analogy it's almost building a laptop right now. There's multiple different components working together, very organic fashion and then you want to make it a, almost a, a native orchestration of all kinds of different components to make it work magically for the scientist.

Professor Louise Serpell

That sounds great. So I guess accuracy is one of the questions that people often ask about AI that sometimes it certainly when I've tried things, I noticed that there are some things are incorrect. How do you ensure that there aren't inaccuracies or errors in the output?

Kexin Huang

Right, yeah, so there's definitely ways to, improve the accuracy of the agents and I would say it's not, still not 100 percent fully accurate across the entire space of the tasks. So that's why the review is still very important. But this indeed recipes to improve it, basically we can usually for a task that's not performing too well, we can just work with the expert that know how to get a task done and we can learn, the agent can basically codify the knowledges of that expert and then the agent basically can imitate the behaviour or the thinking process of experts to get the work done.

And once it's codified, then anyone can basically benefit from it as well. So it's also a democratisation tool where we can codify the best knowledge from the best expert and then distribute it to every AD researchers.

Professor Louise Serpell

Okay. What's your personal hope that somebody will find the answer to?

Kexin Huang

Yes, definitely, I think AD is still, there is still also open question about mechanism about even the cures or treatments. So hopefully this process can at least accelerate the drug discovery process for the AD. Yes, maybe a treatment could be coming from sessions from the Biomni AD that would be the most impactful thing we can ever do yeah.

Professor Louise Serpell

Thank you so much Kexin. So I'm going to turn to Jonathan now. Your team came at this from a different angle. Parthenon has been described as a virtual wet lab with AI co scientist called Athena. That's quite a vivid image for anyone who spent any time at a bench. So let's start with what's under the hood? How does it work exactly and where did the idea of a virtual wet lab come from?

Jonathan Hoover

Yes, great question. So I think there are are two layers to this. There's the models that power the system, and then there's the virtual wet lab piece of it itself. I'll break them down for us, but at a high level, our goal here with these models is to essentially enable researchers to ask if I have cells in a specific cell type and I want to move them to another cell type, what are the perturbations or stimulations or anything you would add to the experiment that would cause that shift?

I think therapeutically this seems is, it tends to map to, if I have cells in a disease state and I want to make them look less diseased, what is the perturbation or sets of perturbations or therapeutic target that could drive that shift? Our models aim to answer that question. So they are built, there's two layers of these models.

There's a cell state model that understands the cell state representation of the cells that we're looking at. And then there is the perturbation model that understands how each of these cells responds to perturbation cell. And so together these help you ask what list of perturbation experiments can I run? What drugs can we use to cause the shift of interest? So that's what the modelling itself does within Parthenon.

The virtual wet lab system makes this model accessible to people who don't necessarily have bioinformatics or model training experience. It includes a discovery engine that allows for analyses and you can ask okay, I want to move cells from a disease associated microglia state to an antigen presenting state. What are the list of perturbations or the set of perturbations that might cause the shift? And it'll pull up a bunch of plots and graphs that make this a lot easier to understand.

Additionally, you can ask if I have a cell in a starting state that is again, let's say go with a disease associated microglia and I add this perturbation to it, what are the differentially expressed genes that I expect to occur that are predicted to occur from this? So it is an analysis tool. On top of that it allows for, we've built in capabilities for generative combinatorial perturbations, essentially can we combine any known set of perturbations?

This could be a gene or a CRISPR knockout, it could be a drug, et cetera, in a way that moves my cells in a direction that wouldn't be possible with one by itself. And this allows for experimental planning. So functionally what we are doing with this system is making it easier for wet lab researchers to shortlist the experiments that would be used to ask their question of interest.

We also have combined it with a system that allows plugging in these agentic AI co scientists by Biomni or other ones out there that can pull on that power that Kexin just described about coming up with or searching literature, coming up with hypotheses about the gene pathways that are involved to also augment what our model is currently capable of doing. And we've built into, with this virtual wet lab, the ability to do model training and searching through the AD Workbench system to find data sets that may have relevant perturbation data sets or relevant cell type data sets to improve the model before you do your experimentation.

So this all creates this loop where an individual researcher can do their, ask their questions, generate data based off of the experimental plan that comes up, put it back into the system, and that will improve the model and kick off this flywheel that provides information back to the community.

Professor Louise Serpell

Okay, so it's not actually replacing the scientists necessarily or the experimental detail, but it's predicting what might happen to allow them to decide exactly what they will refine their experiments to do?

Jonathan Hoover

Yes, exactly. So a lot of these experiments are prohibitively expensive with the CRISPR screens, large scale drug screens. And so shortlisting those candidates better understanding which ones will drive cells to a state of interest that is therapeutically relevant, saves money and increases the return on investment. There are also non therapeutic utilities here, for example, a lot of cell models don't accurately, they're great cell models than the model, the Alzheimer's disease, but they don't perfectly recapture what is happening in the brain.

And with this tool you can start to shift, you can ask what are the perturbations, what are the functionally Yamanaka factors within the context of what we're looking at that would move the cell models to look and more accurately recapitulate human biology. So yeah, it's a partner for a scientist, not a replacement.

Professor Louise Serpell

So before I ask you a load of biology questions, maybe I could just ask you, Jonathan, what your background is and where you came from, how you got to being in this position with Parthenon?

Jonathan Hoover

So I have a winding path. I started in a biology undergrad, then switched to a neuro major, graduated with neuro, did a couple of years in lung transplant immunology lab, then a cancer immunology lab. From there I was "Okay, I don't love being in the wet lab, but I love science." Went and got a master's in statistics at the University of Edinburgh. From there I moved to Genentech, where I was doing single cell RNA sequencing work and comparing a bunch of different technologies. And I have since, for about eight or nine months now, been at Prima Mente working with them on cell state, these virtual cell models, how to build them to accurately allow us to predict changes in cell state with perturbation data.

Professor Louise Serpell

Good, well then I have permission I think to ask you biology questions.

Jonathan Hoover

I'll do my best. It's been a while since I've been in the lab, but yes.

Professor Louise Serpell

Well it's more about the cell models and I'm trying to think about the sorts of experiments we might do in the lab. For example, we might want to use iPSC that have been created from a donor who has a particular mutation. And then to try and examine what changes there are in the pathological proteins associated with Alzheimer's disease, maybe Alzheimer's, maybe amyloid beta and tau. But then to question what perturbations we could make to those cells. And I'm wondering if that's the experiment you are thinking about being able to predict and then guide.

Jonathan Hoover

Yes, 100%. So the way that we define cell state currently is with transcriptomics. We intend to expand this to epigenomics and other functional readouts as well. But in the case of what you're talking, we work a lot with iPSC cell lines, particularly in microglial astrocytes. We are expanding into cultures and tricultures with neurons as well to see how they interact under varied cellular context and perturbation to change there. But yes, you can ask the question, given a genetic background and a perturbation that I want to apply to it, what is the expected change that I would have in the state right now that's transcriptomics, but that could be as we, it increase the data sets that move into this that could be epigenetics, could be proteomics, et cetera.

And this will help one, understand how that perturbation in itself is driving the change. But as you partner with these AI co scientists, maybe that gene is within a pathway that would expand that you can find other targets that you could perturb as well that would have a similar function and that could expand the understanding of the gene regulatory networks that are involved with causing the cell state to change of interest.

Professor Louise Serpell

So in terms of the types of cells that you are using, would you, I remember seeing that it's mostly astrocytes and microglia that are mentioned. But you are including neurons and other cell types in your data set.

Jonathan Hoover

Yes, so the proof of concept that came out at the time of the competition was microglia and astrocytes. But we are designing experiments right now that are all three cells, both monocultures and tri cultures across all three of them. So that we can expand that, yeah, that data corpus.

Professor Louise Serpell

Organoids and mini brains and. ..

Jonathan Hoover

Yes, we are not sure yet if we want to do organoids. They're a very difficult system to get consistent data out of. But it is something that is 100 percent under consideration.

Professor Louise Serpell

Yes, and then we're extrapolating to human networks.

Jonathan Hoover

Yeah we're also very interested in xenograft data as well. So you culture the iPSC cells, do the perturbation of interest and then transplant those into mice with a disease background, et cetera.

Professor Louise Serpell

Yes, that's been very useful I think hasn't it? I want to ask you one more biology question, which is I've understood then from what you've said, that it's mostly about the omics data that you've got in. So the sorts of experiments we would do would be asking questions about the effects on particular parts of the cells, particular organelles mitochondria and lysosomes. So at this point it sounds you are not at that stage yet where you can consider what might happen in a particular part of the cell. It's more of a whole cell question.

Jonathan Hoover

Yeah at the moment, if we move into more spatial imaging, that will be within the realm of what we can look into. We don't have that data at the moment, but it is of interest to us. We can ask at a high level, what is the phagocytic activity that's involved? Like you could do those types of assays and take those readouts from and include them in our representation of cell state. And that is on our path for what we're intending to do.

Professor Louise Serpell

And from the data you'd get out from the omics data, the proteomics or genomics, you might expect to see particular organelles being highlighted in terms of the pathway.

Jonathan Hoover

I must admit this part of the biology I'm not quite as familiar with, but from my understanding, at a high level, yes you could pick out markers that would be indicated in those specific pieces of the biology.

Professor Louise Serpell

Yes, I think that's right. That they would be pointing towards a particular pathway within the cell that might trigger the scientists to think, well we'll go and focus on the mitochondria or focus on the nucleus or nucleolus or something. Thank you. Really helpful. So the judges singled out your platform for being deeply foundational and for letting researchers test therapeutic hypotheses in silico. For someone who designs experiments for a living, what does that change about how they would plan their next project?

Jonathan Hoover

I think that there, this is a tool that powers discovery. So I think really what is changing here is how to come to a hypothesis. This is a hypothesis generation tool. And then with the Athena harness that connects to these AI co scientists, it helps narrow down the exact experimental plan, but it's in a conversation. It's not the scientist is oh, my hands, hands off, see what happens here. It is refining the potential experiments that can be done generating new hypotheses that would not have been thought of because either the data was not clear in the literature or there's just a small piece of the literature that just they hadn't seen.

So it essentially expands where they can look and then narrows down on which ones are more likely to generate an output or have a higher return on investment for them.

Professor Louise Serpell

Okay, and Athena is a co scientist, so it's similar to what Biomni does, you know?

Jonathan Hoover

So Athena is our system to connect co scientists. We are developing our own internal things, but you could plug and play whichever works best for you. For example, you could plug and play Biomni in, you could plug and play NVIDIA's models. You could put in yeah, whatever, whoever we end up partnering with for this. And for any open source models that are out there, those could be included. So it would be powered by the models themselves and then connecting them to our system so that it, so that our models on cell state and perturbation embedding can inform those LLM based models and connect to get a better understanding of what's happening in the field.

Professor Louise Serpell

So before I move on, I'll ask you a similar question to what I asked Kexin. What would you hope to see with the use of this model personally?

Jonathan Hoover

I think there are two things that I would to see out of this. One is more research. I would love if people could make better cell models that accurately represent microglia in the brain. I think that would be phenomenal for everyone. I think the other piece is it would be incredibly exciting to find a perturbation or a small, a CRISPR perturbation or a small molecule that is acting in these pathways that genuinely does shift cells from a more disease state to a less disease state, preferably one that has already been shown to work in the clinic, so that it could be tested more easily.

But yeah, that would be great. Obviously the post discovery piece of all of this, the clinical trials is very difficult and, but it would be great to to see a translation there.

Professor Louise Serpell

Okay. Thank you very much Jonathan. So question for both Jonathan and Kexin. Both of you have built tools that are pitched as collaborators rather than replacements, which comes as a huge relief as a scientist and probably for most of our listeners who are pursuing this career. Where do you think the line genuinely sits between a useful AI partner and overreach, and what do you think a researcher should absolutely not delegate? So maybe I'll go back to you first Kexin.

Kexin Huang

Yes, I think this is a great question. So I'll say it's definitely AI, right now's more a research assistant in contrast to research overlord. As we design a platform humans still need to provide the prompt to that agent to perform a task. So with all the prompt, it cannot do too much, right? So that's why we always design a platform and make it an AI partner instead of an overreach.

Yes, and for the task, that researcher absolutely not delegate, I would say actually I think it's okay to delegate as much as possible, but it's just researchers need to review the answers and the traces because indeed the agent is an extremely useful platform. It can dramatically increase the productivity and they can do all kinds of different tasks that scientists is performing day to day. But the answer still requires review and validation from the humans as well.

Professor Louise Serpell

Thank you very much. And Jonathan?

Jonathan Hoover

Yes, I think I'm aligned with Kexin here. AI is incredibly useful. I think as every day the line for where you would not delegate shifts and changes as the hallucination guardrails get better and the models get better, you can delegate more or okay, maybe delegate is not the right word, but you can trust the outcomes more. I will say there are a lot of things that you can and should delegate. It is useful tool that essentially allows you to free up your own thinking space. But that means that you have to use that space, your own, your own reasoning in the right places.

So I will let these agents do simple math for me, write code for me that is fairly simple implementation, but high level systems, I'm the one who's reviewing all of these things. So yeah, it depends on the scale that you are doing this work. But at the end of the day, a human should currently still be in the loop for review for all of these things.

Professor Louise Serpell

Okay, thank you very much. We are nearly out of time, but before we wrap up, I'd to ask a final question. So what would you to see happen, perhaps an action that each dementia researcher might take after listening to this podcast today? And I'll start with Kexin first.

Kexin Huang

Yes, I think so agent is definitely going to be the new way how scientists are going to work in, in a couple of years. So I'll definitely encourage scientists and researchers to start experimenting and then also embrace the technology and then also try all kinds of different tasks in the platform while looking at the result closely and validate and review it as well.

Professor Louise Serpell

Okay that sounds brilliant. And Jonathan?

Jonathan Hoover

Take a look at the AD Workbench. I think there's a lot on there that is helpful and useful. And then also understanding the gaps for all of us of what is missing is also very, very helpful for all of us. For anyone, I think it's very important to become literate in AI tools at the moment. I think we're at a place where it's similar to the early two thousands where you need to learn how to use Google. This is how research is going to be done. And it does not have to just be for programmers, you can build an agent as a wet lab scientist that does research for you, that will help you send emails, all of these things.

It's a useful tool to learn and understand. And I think that if it is not done now, it is easy to fall behind. And so that's something that I would suggest people get on now.

Professor Louise Serpell

Yes, I definitely feel excited to try it out myself now. So yes, to become much more literate with it. Thank you. And Bose, what do you think you'd really to see happen from the launch and from this prize now?

Dr Niranjan Bose

Join us on this journey. I think we need ideas from all different disciplines. We need feedback on the tools and the data sets. And by joining us on this journey, you're going to make the global research ecosystem a lot stronger, a lot better. And I'm optimistic we will shorten the time it takes from studies to data to insights.

Professor Louise Serpell

Yes, it's fantastic. It sounds so collaborative and really exciting. I guess one of the things that maybe the field is going to have to think about is exactly how people are judged in terms of their outputs and the success in the field because we all want to collaborate and share our knowledge. And what we all want really in the end is a treatment and diagnosis, early diagnosis, being able to identify the right people and treat them in using personal medicine. And the problem at the moment that I see in terms of the scientific field is that we are all judged by our papers and our research grants.

And if we are all going to collaborate, this will make for a better world, but perhaps a more complicated way of ranking and judging people. So that would be a very interesting future for us all I think. Everyone will have to embrace that perhaps. So thank you all so much. I wanted to just reflect a little bit here that from this conversation it sounds that the dementia research bottleneck is really shifting, but we used to talk about not having enough data and now we have loads of data and we're really moving towards knowing how to actually treat that data and make real strides forward in terms of understanding how to use it and how to then iteratively plan our new experiments and ask questions about it.

I think that the tools are very clearly being provided to not replace researchers, but to work with them and help them with their research. So that's really exciting and I think there's been some real insights into how somebody who might want to contribute to this whole endeavour might be able to put forward their ideas, contribute data, and ask the questions that really need to be asked to understand Alzheimer's disease and dementia.

So finally, I'd just to thank you very much both Jonathan and Kexin for joining me today. You can find out more about the Alzheimer's Disease Data Initiative, AD Workbench and the prize at the AD Data Initiative website. And as always, you'll find all of the resources we've mentioned on the show notes for this episode at dementiaresearcher. nihr. ac. uk. Until next time, I'm Louise Serpell and you've been listening to the Dementia Researcher Podcast. Thank you very much. Thank you.

Kexin Huang

Thank you.

Jonathan Hoover

Thank you, Louise.

Narrator

The Dementia Researcher Podcast was brought to you by University College London with generous funding from the UK National Institute for Health, Alzheimer's Research UK, Alzheimer's Society, Alzheimer's Association, and Race Against Dementia. Please subscribe, leave us a review and register on our website for full access to all our great resources. Dementiaresearcher. nihr. ac. uk.